JP2022512402A - Optimization of glycemic datasets for improved hypoglycemia prediction based on the incorporation of machine learning implementations - Google Patents

Abstract

The present invention relates to a method for data set expansion for improved hypoglycemic prediction based on classifier uptake, the process of providing a raw data set for a subject, wherein the data set is located. The steps to be provided, including multiple BG values obtained at a given sampling rate, and the time stamps associated with those values over multiple days N, and the evaluation block value (eHH) of the rolling scheme as input X. A sliding window containing a BG value for a given past period Tp, including the step of performing data conversion by temporal binning to create the corresponding predicted value (pHH) as the output Y. Is created as an indicator I indicating whether the BG value at a given future time Tf is below a given threshold indicating a hypoglycemic condition. [Selection diagram] FIG. 13

Description

The disclosure generally relates to systems and methods for assisting patients and healthcare professionals in the management of insulin treatment for diabetes. In certain aspects, the invention relates to methods for higher data resolution optimized for incorporating machine learning (ML) implementations.

Diabetes mellitus (DM) is impaired insulin secretion leading to hyperglycemia and varying degrees of peripheral insulin resistance. Type 2 diabetes mellitus is characterized by a progressive obstruction of normal physiological insulin secretion. In healthy individuals, pancreatic β-cells continuously secrete basal insulin, maintaining steady-state glucose levels between meals for extended periods of time. In healthy individuals, insulin is also rapidly released during the initial first-stage spikes corresponding to the diet, followed by long-term insulin secretion that returns to basal levels after 2-3 hours. Years of uncontrolled hyperglycemia can lead to multiple health complications. Diabetes mellitus is one of the leading causes of early morbidity and mortality worldwide.

Effective control of blood glucose / plasma glucose (BG) can prevent or delay many of these complications, but once established, they may be irreversible. Therefore, achieving good glycemic control in efforts to prevent diabetic complications is a major goal in the treatment of type 1 and type 2 diabetes. In particular, frequent changes in insulin dose regulation are key to help stabilize the patient's blood glucose levels (Bergenstal et al., "Can a Tool that Automatic Titration be a Key to Diabetes Management?" and Thera. 2012; 14 (8) 675-682). To administer the insulin drug treatment regimen, smart titrators have been developed that use adjustable process sizes, as well as physiological parameter estimates and predetermined fasting blood glucose target values. The optimal initiation and titration method for long-acting basal insulin remains to be determined. However, the evidence is that many patients often do not receive adequately titrated insulin doses to reach their glucose control target levels (remaining suboptimal doses and failing to reach therapeutic goals). (Holman et al., "10-ear follow-up of intensive glucose control in type 2 diabetes," N. Engl. J. Med. 2008; 359: 1577-1589).

One of the major problems with insulin regimens is the lack of patient autonomy and empowerment. Patients often have to visit the clinic to calculate a new titration. If the clinic has to titrate a patient's insulin dose, there is a natural limit to the frequency of titration dose changes. Self-titration regimens can promote patient empowerment and allow them to be more involved in treatment, resulting in improved glycemic control (Khunti et al., “Self-titration of insulin in the management”). of people with type 2 diabetes: a practical solution to impedance management care, "Diabetes, Obes., And Metabol. 2012; 15" (15). Patients who play an active role in diabetes management and insulin titration are responsible for their own self-care, strongly believe that their actions can affect their disease, and can lead to better treatment results. There is a possibility (Norris et al., "Self-management education for adults with type 2 diabetes: a meta-analysis on the effect of glycemic control "Effects of self-management training in type 2 diabetes: a randomized, prospective trial," Diabet.Med.2007; 24:. ". Patient empowerment: results of a randomized controlled trial" 415-23, Anderson et al, Diabetes Care 1995; 18: 943-9). In addition, if the patient manages his or her own titration, the frequency of titration will increase, thereby increasing the likelihood that the patient will achieve the desired blood glucose level.

However, with a more aggressive titration approach, the risk of hypoglycemic events (“hypoglycemia”) is higher, and with multiple daily injection (MDI) -based titration regimens, the risk is even higher. In contrast, some solutions for short-term hypoglycemic prediction (STHP), such as the “Evolution of a New Measure of Diabetes” from Kovatchev et al. (TypeZero & University of Virginia group), 11), November 2006, Sparacino et al. (Cobelli Lab in University of Padova) of "Glucose Concentration can be Predicted Ahead in Time From Continuous Glucose Monitoring Sensor Time-Series", IEEE Transactions on Biomedical Engineering, Vol. 54 (5) May 2007, Franc et al. (Volunts with Sanofi)'s "Real-life application and validation of flexible insulin-therapeutic algorithm" In December 2009, 35 (6): 463-8, and Sudharsan et al. (WellDoc) (LTHP 24-hours ahead literature comparison) "Hypoglycemia Prediction Using Machine Learning Models for Patients with Type 2 Diabetes", Journal of Diabetes Science and Technology 2015, Vol. 9 (1) 86-90 has been proposed.

To address this issue, US2008 / 0154513 discloses methods, systems, and computer program products related to maintaining optimal control of diabetes, and blood glucose readings collected by a blood glucose self-monitoring (SMBG) device. Based on, it is intended to predict patterns of hypoglycemia, hyperglycemia, increased glucose variability, and inadequate or excessive testing over the coming period. Methods for identifying and / or predicting a user's pattern of hyperglycemia include the steps of acquiring multiple SMBG data points, classifying SMBG data points within a given duration, and each period. It comprises the step of evaluating the glucose level of the blood glucose level and the step of showing the risk of hyperglycemia over a subsequent period based on the above evaluation. The evaluation includes a step of determining the individual deviation for hyperglycemia based on the glucose value, a step of determining the combined probability of each of the above periods based on the individual deviation and the absolute deviation, and the above-mentioned combined probability of each period in advance. A step of comparing with a set threshold may be included. The period may include dividing the 24-hour day into time bins with a predetermined duration.

In order to address the above-mentioned problems and better reduce the risk of hypoglycemia, an object of the present invention provides methods and systems for predicting future hypoglycemia and improving the ability to reduce the current recommended dose. And this allows for a more accurate titration regimen, thereby enabling the treatment of type 2 diabetes. A particular object of the invention is to provide a method for dataset optimization that allows for improved classifier uptake and improved hypoglycemia prediction based on machine learning algorithms. Such methods need to use a transparent and constrained approach to make them more suitable for use in the dosing guidance system with the approval of the authorities.

US2008 / 0154513

"Can a Tool that Insulin Titration be a Key to Diabetes Management?" "10-year follow-up of glucose glucose control in type 2 diabetes" "Self-titration of insulin in the management of people with type 2 diabetes: a practical solution to impact management in primary" "Self-management education for adults with type 2 diabetes: a meta-analysis on the effect of glysmic control" "Effects of self-management training in type 2 diabetes: a randomized, positive trial" "Patient empowerment: results of a randomized controlled trial" "Evaluation of a New Measurement of Blood Glucose Variation in Diabetes" "Glucose Concentration can be Predicted Ahead in Time From Glucose Monitoring Sensor Time-Series" "Real-life application and validation of flexible insulin-therapy algorithms in type 1 diabetes patients" "Hypoglycemia Prediction Machine Learning Machines for Patients with Type 2 Diabetes"

Means for Solving the Problems In the disclosure of the present invention, one or more of the above objects are addressed, or the objectives apparent from the description of the exemplary embodiments as well as the following disclosure are addressed. The morphology and aspects are described.

A first aspect of the invention is a method for dataset optimization for improved hypoglycemic prediction based on classifier uptake, wherein the method provides a raw dataset for the subject. The process of providing and the evaluation block that the dataset comprises, including multiple BG values acquired at a given sampling rate, and a time stamp associated with those values over multiple days N. A step of performing data conversion by temporal binning of a rolling scheme with a value (eHH) as input X to create a corresponding predicted value (pHH) as output Y, wherein X is a given past period. Created as a sliding window containing a BG value for Tp, an indicator I indicating whether Y is below a given threshold indicating a hypoglycemic condition at a given future time Tf. A method is provided, which is created as.

In general, the predictive model depends on the data being trained. The above method allows the same amount of data to be utilized in a more efficient and better way that fits and fits machine learning algorithms such as random forest (RF) classifiers accordingly.

In contrast, previous attempts aimed at predicting patterns of hypoglycemia, as disclosed in US2008 / 0154513, were simple temporal binning of BG data, followed by organized data. It has relied on traditional mathematical analysis.

Data conversion can be performed over at least two different past periods Tp. Tf may correspond to Tp, for example, the 15 minute predicted value is based on the 15 minute BG value.

In an exemplary embodiment, the data conversion step is performed after the step of performing data expansion by temporal binning of the daily BG value rolling scheme to the M-day evaluation block, where M is 2 or greater. Yes, and less than N for multiple days.

Such data expansion is usually based on an M-day insulin titration regimen, eg, 3 days with the same insulin dose before modification, if the raw data set obtained is usually of a given basal insulin. Used for basal insulin titration as indicated in the instructions for use. For bolus insulin-based datasets, M = 1 would be appropriate. In fact, when M = 1, no actual expansion is done.

In an exemplary embodiment, the step of providing a raw data set performs data preparation with resampling corresponding to the nominal sampling rate and creation of interpolated BG values to replace the missing BG values. It is done before the process of

A further aspect of the invention provides a method for training a classifier, which captures a classifier-optimized dataset and trains the classifier based on the captured dataset, as described above. Includes the step of providing a personalized dataset. The classifier may be a random forest classifier.

In a further aspect of the invention, a method for predicting future BG values, a step of obtaining a series of evaluations of BG values from a subject, and a series of BG values in a classifier trained as described above. A method is provided that includes a step of incorporating the evaluation of the above and a step of providing a predicted BG value. The classifier-trained dataset may be obtained from the same subject as the series of assessments of BG values. A series of assessments of BG values can be obtained by continuous blood glucose monitoring (CGM), eg, generating BG values every 5 minutes.

A still further aspect of the invention is a computer processing system for performing temporal optimization of a data set from a subject, wherein the computer system comprises one or more processors, a memory, and a memory. However, there is provided a computer processing system that includes instructions and, when the instructions are executed by one or more processors, implements the methods defined above according to different aspects of the invention.

Certain exemplary embodiments use the same amount of data, but time optimization and expansion of the data in a more expanded and smarter fitting way is provided by performing the following steps: Will be done.
(1) Handling of missing data: 5 minutes of resampling using a spline interpolation solution: The data size increases with missing data to meet the data quality processing requirements of data preparation with part of the software code. ..
(2) Time binning evaluation limit history (eHH) of the rolling scheme: a series of CGM measurements nested within the clinically derived interval of the 3-day study block or the evaluation limit history (eHH) 3 days ago. Three-day block binning with a time-optimized daily rolling scheme as opposed to the standard sequential scheme for binning.
(3) Predictive Limit History (pHH) of hypoglycemia by temporal binning of rolling schemes: 15 minutes, 30 minutes, and 60 minutes ahead based on some future time interval, corresponding previous retroactive time interval. A software program that repeatedly predicts hypoglycemia at the predicted limit (PH) of, or the predicted limit history (pHH) 15 minutes, 30 minutes, and 60 minutes ago, respectively. In each step, pHH = PH prediction is performed every 5 minutes in the rolling scheme as opposed to the sequential scheme.

Combining all three steps increases the size and depth of the original unprocessed BG dataset. Therefore, the processed dataset converted by the three-step technique directly and quickly to the ML classifier format achieves not only significantly larger size, but also depth and operational capture potential. Raw or raw datasets cannot be easily or immediately populated or fed into the ML classifier format with the same efficiency.

Taken together, spline interpolation of missing data by temporal binning of the rolling scheme of the interval between the evaluation limit history and the prediction limit history is highly sensitive (accurate prediction of hypoglycemic events) and highly specific (non-hypoglycemic events). Accurate Prediction) provides optimization of CGM resolution data for more accurate prediction of hypoglycemia.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
An exemplary data preparation module according to an embodiment of the present disclosure is shown. An exemplary data conversion module according to an embodiment of the present disclosure is shown. An exemplary pointer look-up table according to an embodiment of the present disclosure is shown. An exemplary temporal binning optimization according to an embodiment of the present disclosure. An exemplary hypoglycemia determination module for different pHH values according to one embodiment of the present disclosure is shown. An exemplary hypoglycemia determination module for different pHH values according to one embodiment of the present disclosure is shown. An exemplary hypoglycemia determination module for different pHH values according to one embodiment of the present disclosure is shown. The preservation of exemplary training results for subsequent ML processing according to one embodiment of the present disclosure is shown. An exemplary implementation of a random forest (RF) classifier according to an embodiment of the present disclosure is shown. An exemplary implementation of a random forest (RF) classifier according to an embodiment of the present disclosure is shown. The results of the RF classifier according to one embodiment of the present disclosure are shown. The results of the RF classifier according to one embodiment of the present disclosure are shown. The results of the RF classifier compared to the results in the literature are shown. The results of the RF classifier compared to the results in the literature are shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown. The embodiments according to one embodiment of the present disclosure are collectively shown.

In the figure, similar structures are identified primarily by similar reference numbers.

The present disclosure relies on the acquisition of a set of training and test data containing information relevant to at least one subject. The dataset (s) are available over time for multiple glucose measurements of the subject obtained over time and for each glucose measurement at multiple glucose measurements to establish a glycemic history. Containing at least a corresponding glycemic time stamp indicating at what point in time each glucose measurement was taken and one or more basal insulin injection histories, the injection history is multiple during all or part of the time course. And for each injection in multiple injections, at least the amount of the corresponding dosing event and the time stamp of the dosing event indicating at what point in the time course each injecting event occurred.

STHP classifier: Data preparation and data transformation of the STHP classifier 15 minutes to up to 60 minutes ahead, then currently experimental, to determine the prediction or detection of adverse events of hypoglycemia or hypoglycemic levels in a short period of time. The predictive limits (PH) of the above and future machine learning methodologies range from self-monitoring of blood glucose (SMBG) at 1 or 2 points per day to flash glucose monitor (FGM) at 15 minute intervals or 5 minute intervals. Up to the continuous glucose monitor (CGM), optimization and adaptation are required to fully capture, adopt, and leverage different temporal resolutions.

In general, the predictive model depends on the data being trained. Therefore, improving data quality or using the same amount of data more efficiently is the heaviest and most valuable. This current solution not only leverages more data with the higher temporal resolution of CGM, but also fits and adapts this data to machine learning algorithms such as random forest (RF) classifiers. We also aim to use it in a smarter and better way. For example, in the space between 12:00 pm and 3:00 pm, with the low resolution of SMBG with an hourly report by Dexcom, it is possible to acquire only three intervals. With CGM high resolution and full data optimization, 25 intervals can be obtained and fed to ML models such as random forest (RF) classifiers.

The current configuration or use of CGM data for the Random Forest classifier algorithm is as follows: For example, in order to predict hypoglycemia for the next 60 minutes (PH = 60 minutes ahead), the past 60 minutes is used as the input prediction limit history (pHH), but the past 3 of the evaluation limit history (eHH) is still used. Limited to within the block of days. If there is no CGM data and only SMBG data, the temporal shift occurs every hour.

For example, in SMBG data, for example, in a three-hour space from 12:00 pm to 3:00 pm, the interval between the three temporal data, 1) the first interval from 12:00 pm to 1:00 pm, and 2) afternoon. There is only a second interval from 1:00 pm to 2:00 pm and 3) a third interval from 2:00 pm to 3:00 pm.

This makes sense within the constraints of other measurement schemes such as SMBG or other equipment, but not in CGM. This lower resolution scheme does not allow the higher resolution data from CGM to be optimized and fully utilized.

Time optimization of CGM adapts 25 temporal data intervals, each 5 minutes apart, within the same space of 3 hours, as constrained by CGM.

From 12:00 pm to 3:00 pm: SMBG low resolution (Dexcom reports hourly): 3 intervals, CGM high resolution (fully optimized): 25 intervals.

In summary, the above is an optimization of temporal binning of predicted limit history (pHH). Therefore, with this temporal data optimization and adaptation, instead of preparing only 3 data intervals for the machine learning random forest classifier, 25 data temporal intervals are prepared and the machine learning random forest classifier is prepared. Prepare for, thereby increasing data availability and training use cases.

Not surprisingly, this full utilization of CGM data can only be seen as the next logical step, and the real improvement lies in the application of higher resolution of CGM data to machine learning algorithms, some of which. For example, the time series ARIMA model is for capturing daily variation, even if there is a clear daily strong seasonal component in the data, which is captured by other functions such as seasonal_decompose from the statusmodels package. Many of the seasonal parameters (288 points per day) do not work.

Without these CGM data optimization and fitting methods and functions, machine learning algorithms such as Random Forest classifiers are not well trained, do not fit, and cannot represent the data for which predictions are being made.

The medical and scientific basis for utilizing the intermediate 5 minute interval is 15 minutes, 30 minutes, or 60 minutes, respectively, as long as the assumptions of temporal linearity, order, and minimum data quality are maintained. Intervals are estimated only for the future, followed by linear shapes in 5-minute increments, 12 pm to 1 pm window and 12:05 pm, except for new data trends that can be captured within the new window. It makes no difference whether you apply the window from minutes to 1:05 pm.

For example, if 12:00 pm to 1:00 pm is missing within the SMBG resolution of a single point every hour instead of the CGM resolution of a single point at intervals of 5 minutes, that data. There is no way to write, except by high-risk extrapolation. The CGM resolution lacks the interval from 12:00 pm to 1:00 pm, but if 12:05 pm to 1:05 pm is available, the CGM is shifted by 5 minutes every 12 pm The duration of 1 hour from 05 minutes to 1:05 pm is the accepted data.

In SMBG resolution, if the interval from 1:00 pm to 2:00 pm is missing, interpolate between the interval from 12:00 pm to 1:00 pm and the interval from 2:00 pm to 3:00 pm Is possible, and there is some risk, but not as much as extrapolation. In the SMBG resolution, if the interval from 2:00 pm to 3:00 pm is missing, it is the same situation as if the interval from 12:00 pm to 1:00 pm is missing, and the missing data is written. Therefore, extrapolation is required. Basically, interval edge cases require extrapolation, while missing data cases or intervals require interpolation. Both are risky, but interpolation is less risky than extrapolation.

The CGM data optimization process takes advantage of higher resolution and, of course, within medical constraints, can instead rely on other 5-minute-shifted 1-hour intervals for this interpolation and extrapolation. Remove the need. For example, if it is missing for more than 20 minutes, the interval from 12:00 pm to 1:00 pm is assumed to be the interval from 12:25 pm to 1:25 pm (12:00 pm to 12:25 pm). It is not a good idea to replace it with a state in which all the intervals up to the minute are missing, basically five intervals are missing). Otherwise, from a medical, scientific, and physiological point of view, they can be replaced, averaged, or interpolated with each other within 20 or 45 minute intervals, which is missing data. A fitting function that can reliably fit or fit machine learning algorithms such as random forest classifiers, even if they are incomplete or corrupted, as long as some thresholds of data quality and linearity are met. Allows the description of this to be much more stringent and demanding thresholds with lower temporal data resolution of MBG and other methodologies and devices, whereas CGM has higher temporal data resolution. It is a loose threshold.

Another way to think about this is with respect to data quality: CGM optimizations that use everything possible (but linearly constrained) have room for data loss or corruption, and machine learning algorithms such as random forest classifiers are sufficient to generate predictions. Data is still available. If the SMBG has only three intervals, the machine learning algorithm of the random forest classifier is interrupted and cannot give a prediction for the next time, even if one interval is missing.

Hereinafter, an exemplary embodiment of the data preparation module in the Jupyter Notebook code will be described. See FIG.

The data preparation module employs the "convertToTS" and "removeNaNdays" functions. The "removeNaNdays" function itself adopts the output lookup table "pointerTable" of another function, which is covered in the process of the data conversion module. Finally, the "interpolateList" function is adopted. See FIG.

More specifically, the following is done:
1. 1. The CGM data of the target person is read. The subject's CGM data is an object type of tabular data frame.
2. 2. The subject's CGM data (if available) removes any "SMPG" or other data label, leaving only the "CGM" data label.
3. 3. When the "convertToTS" function is adopted, the subject's CGM data (usually in tabular form) is converted into a time series object for further data preparation.
4. Adopting the Pandas time series native resampling function with the mean value of the subject's CGM time series object data, including only the days with at least some CGM data, is "5-T" or 5 Further prepared by resampling into a minute bin. In the absence of missing data, this step yields the same dataset, but stacks neatly for data analysis. For example, at 12:01:43 pm at 85 mg / dL, it is 12:00 pm at the same 85 mg / dL. Further, 12:06:21 pm at 92 mg / dL becomes 12:05 pm at the same 92 mg / dL. If there is missing data, this resampling step must first substantially increase the original raw dataset to a larger dataset that has been processed and then convert it to the actual value in subsequent steps. Generates new missing data or NaN. However, first it is necessary to remove any complete NaN days. In clinical studies, a full NaN day is basically the period between baseline and follow-up days. The resampling process unfortunately adds an unnecessary missing NaN day non-observation period that needs to be programmatically removed, as both baseline and follow-up timestamps are in one data object. do. This is achieved in the next step.
5. The "removeNaNdays" function is adopted.
Input: Subject's CGM [Time Series] Object Data Type Processing: Scanning and removing completely missing NaN days Rationale: Interpolating the entire day between days is also a risk. A much lower risk is to interpolate the CGM values during the same day, which will be the next and final step of data preparation.
Output: Target CGM [List] object data type. [Time Series] Object data type is gone!
This function adopts the "pointerTable" function described in the process of the data conversion module.
6. Employing the "interpolateList" function, this list of erased processed CGM values will eventually write any NaN or missing data within the day range when at least some CGM is available. Interpolated by spline interpolation.

Next, the data conversion module adopts the output of the "pointerTable" function of the CGM look-up table of 288 points per day. See FIG.

More specifically, the following is done:
1 The "pointerTable" function only creates a look-up table that cross-references the 288-point CGM as an ID once.
2. 2. When the "pointerTable" function is adopted, the list of CGMs is assigned 288 cross-referenced IDs to adjust at what point or time stamp of the day a particular value is located.

CGM Pointer Table Lookup Submodule From a medical and scientific point of view, CGM data points are used to determine and confirm fasting plasma glucose (FPG) in the morning or evening afternoon, especially at night. It is important to know if it is associated with a time zone and a morning time zone. A pointer for a single day to still get such information, without a time series object, just by having a list object of CGM values by cross-referencing the 288 IDs of a typical CGM day. I devised a look-up table.

By utilizing the 288 IDs of a typical CGM day pointer table, the timestamp component can be stripped to leave only a list of CGM values. This list of CGM values can then be fed and populated into the ML classifier format algorithm. Unfortunately, the time series object itself cannot be supplied to the ML classifier format algorithm. Therefore, a cross-reference with the CGM 288-point ID table is required.

To retain time-of-day or time information (eg, id = 10 of the 288 CGM points of the day corresponds to 0:50 am or 12:50 am). Create a pointer lookup table for the CGM process with 288/5 points daily. See FIG.

For the top (left figure), the pointer table id = 9 corresponds to the actual time point at 12:45 am, and for the bottom (right figure), the pointer table id = 287 is 23:55 pm. Or it corresponds to 11:55 pm.

Therefore, such a pointer lookup table iterates over a list of CGM values (which may include several days, eg 14-16 days) and CGM at any time of the day without data at the time available. It becomes possible to understand what the value is pointing to. Therefore, since the pointer index 0 corresponds to the new day at 12:00 am, it is possible to divide the long list of CGM values into daily chunks.

Since pointer ID = 0 indicates a new day or the next day, the total list of CGM values can stop inputting to the stand-alone list before that day and start a new stand-alone list of CGM values for the next day. In addition, the algorithm adds only a full day, including all 288 points. Days with less than 288 points will not be added as a full day. For example, in most clinical or realistic clinical trials of users or patients, the first and last days or days usually have less than 288 points. It is best not to utilize such data because it is difficult to extrapolate, interpolate, or write missing data for the corner edge caps of such data. Finally, the algorithm also handles end cases. Otherwise, the final day will not be added properly, as confirmed by the exam. As a result, the total list of CGM values is now binned into daily chunks or blocks.

Therefore, the pointerTable is called only in two places in the STHP classifier codebase.
1. 1. A "removeNaNdays" feature was employed to identify and specify the days that are completely missing or the NaN days for subsequent removal.
2. 2. Mainly, we adopted the process processing (statement in the case of a loop) of the data conversion module whose task is to create the evaluation limit history (eHH) of the block for 3 days from the block of a single day.
Input: Clean list of CGM values Processing: Cross-reference output with pointerTable output of "pointerTable" function: First binning to daily list of CGM values (288 points per day or chunk by day)

The following describes a data optimization module that provides higher temporal resolution optimization of CGM by temporal binning of rolling schemes. Fit for importing into machine learning random forest classifiers

Evaluation Limit History (eHH) -Optimization of temporal binning. See FIG.
Input: Daily list of CGM values. However, it has not yet been binned into a 3-day mass or block.
Treatment: Utilization of the first step of temporal binning of the rolling scheme Output: These daily chunks can then be binned into 3 day chunks or blocks.
Evidence: daily and 3 days based on medical and scientific considerations and guidelines for patient physiological regulation periods, as well as manageable input considerations for model training periods to feed random forest classifiers. Binning into a lump of.
1. 1. The main part of the loop deals with converting a daily history chunk into a three-day limit history (HH) chunk.
2. 2. Adopting the "reduce" function from the "functools" package simply converts, reduces, or flattens the resulting list of lists into a single running list. Each list of times represents a clinically necessary three-day observation or assessment rather than a single day.

So far, CGM data has only one substantial opportunity to increase with a 5-minute resampling function. All of the interpolation function's actions were to write missing NaNs that had already been augmented or extended by a 5-minute resampling step. Therefore, the interpolation function cannot grow or expand the data. Similarly, binning to the daily chunks is set up simply to indicate the number of days available in the subject's CGM data. No overall data expansion is done in this process. Again, the first substantial opportunity to grow the dataset was a 5T or 5 minute resampling step.

However, in this step of binning into blocks for 3 days, there is a second substantial opportunity to grow and expand CGM data.

Typical 3-day block binning to 12-day available total blocks: 4 intervals achieved.

The above typical scheme is meaningful for SMBG or other device data and requires significant recalibration and calculations between each study block for 3 days. However, this makes little sense for CGM data, which requires only calibration (1-2) within the same day and can perform calculations daily. Therefore, there is no point in missing a block for 3 days, such as from the 2nd to the 4th day. The medical, scientific, and data science assumptions also apply to the complete data optimization of this rolling scheme with the higher temporal resolution data of CGM. These assumptions do not apply to SMBG and other device data, so typical schemes are used. However, this typical scheme is suboptimal for implementing CGM, especially for incorporating ML classifiers. Of course, the problem of rolling schemes is further solved and fitted in detail so that it can be quickly adopted by the ML method, from Random Forest (RF) to Support Vector Machines (SVMs) to k-nearest neighbors (KNN). Will be done.

Accordingly, the following optimized and more data collection methods for binning blocks for 3 days are provided.

Optimized 3-day block binning to total blocks available for 12 days:

Ten intervals were achieved with this optimized scheme. Basically, a total of n-3 pieces is included.
Optimization of temporal binning of hypoglycemic predictive limit history (pHH):
Input: A 3-day chunk or block of evaluation limit history (eHH).
Rationale: This setup avoids time pitfalls and bleeding errors into the next 3-day clinical evaluation period. Properly packaged for ML analysis.
Treatment: Utilization of the second step of temporal binning of the rolling scheme Output: Predicted limit history (pHH) is nested within a 3-day chunk or block evaluation HH (eHH). This is essential for setting up boundaries and boundaries that are contoured for machine learning (ML) and can also comply with the patient's physiological adjustments or alignments. In this second innovative process, this is a third substantive opportunity to increase the input data. Therefore, the original raw input data is augmented or purified into processed and purified input data ready for ML classifier format capture, modeling, training, and testing in three substantive steps. It has been expanded.

See FIG. 5 for pHH = PH = 60 minutes.

This final data optimization step is to modularize these predicted limit histories (pHH) of the ML classifier inputs separately from the data preparation, transformation, and data conformance by creating the evaluation limit history (eHH). Now, only this determination of hypoglycemia varies between different implementations, from pHH = PH = 15 to pHH = PH = 30 minutes, pHH = PH = 60 minutes.

See FIG. 6 for pHH = PH = 30 minutes and FIG. 7 for pHH = PH = 15 minutes.

So far, the exemplary embodiment is behind converting raw, unprocessed CGM data into data that has been extended three times, time-optimized, purified, processed, and incorporated into the ML. It covers computer-processed calculations in and can be fed into a random forest (RF) classifier model.

In the final data section, which focuses on the generation and storage of the training-test X-y set (see Figure 8), both the independent variable (X) and the dependent variable (y) are the training-test split dataset. It is saved with the part of. These final datasets for this particular pHH = PH = 60 minutes are then validated in the test code section. See below.

After these final datasets have been saved, the actual STHP RF classifier model can be run and created using the final data inputs.

ｐＨＨ＝ＰＨ＝１５スライディングウィンドウ６つ。
入力：ブロック１のｅＨＨ：
０：１日目：［１５８、３３５、１４６、３７１、１０４、１７０、１０９、２９０、１２７、１５１、２３１、３７６］
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ１＝［１５８、３３５、１４６、３７１、１０４、１７０］
Ｘ１＝［１５８、３３５、１４６］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ１＝０～１７０＞７０＝０、１７０ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
したがって、Ｘ１がＸ（または入力、過去のＣＧＭのＢＧ値）に追加または付加されることになり、Ｙ１がＹ（出力、低血糖／低血糖なしのバイナリ分類子、オン／オフ）に追加または付加されることになる。
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ２＝［３３５、１４６、３７１、１０４、１７０、１０９］
Ｘ２＝［３３５、１４６、３７１］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ２＝０～１０９＞７０＝０、１０９ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
ここまでのＸおよびＹは以下のとおり。
Ｘ＝［［１５８、３３５、１４６］、～Ｘ［０］
［３３５、１４６、３７１］］～Ｘ［１］
Ｙ＝［０、０］～Ｙｓ［０］、Ｙｓ［１］
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ３＝［１４６、３７１、１０４、１７０、１０９、２９０］
Ｘ３＝［１４６、３７１、１０４］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ３＝０～２９０＞７０＝０、２９０ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
ここまでのＸおよびＹは以下のとおり。
Ｘ＝［［１５８、３３５、１４６］、～Ｘ［０］
［３３５、１４６、３７１］、～Ｘ［１］
［１４６、３７１、１０４］］～Ｘ［２］
Ｙ＝［０、０、０］～Ｙ［０］、Ｙ［１］、Ｙ［２］
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ４＝［３７１、１０４、１７０、１０９、２９０、１２７］
Ｘ４＝［３７１、１０４、１７０］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ４＝０～１２７＞７０＝０、１２７ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
ここまでのＸおよびＹは以下のとおり。
Ｘ＝［［１５８、３３５、１４６］、～Ｘ［０］
［３３５、１４６、３７１］、～Ｘ［１］
［１４６、３７１、１０４］、～Ｘ［２］
［３７１、１０４、１７０］］～Ｘ［３］
Ｙ＝［０、０、０、０］～Ｙｓ［０］、Ｙ［１］、Ｙ［２］、Ｙ［３］
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ５＝［１０４、１７０、１０９、２９０、１２７、１５１］
Ｘ５＝［１０４、１７０、１０９］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ５＝０～１５１＞７０＝０、１５１ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
ここまでのＸおよびＹは以下のとおり。
Ｘ＝［［１５８、３３５、１４６］、～Ｘ［０］
［３３５、１４６、３７１］、～Ｘ［１］
［１４６、３７１、１０４］、～Ｘ［２］
［３７１、１０４、１７０］、～Ｘ［３］
［１０４、１７０、１０９］］～Ｘ［４］
Ｙ＝［０、０、０、０、０］～Ｙ［０］、Ｙ［１］、Ｙ［２］、Ｙ［３］、Ｙ［４］
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ６＝［１７０、１０９、２９０、１２７、１５１、２３１］
Ｘ６＝［１７０、１０９、２９０］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ６＝０～２３１＞７０＝０、２３１ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
ここまでのＸおよびＹは以下のとおり。
Ｘ＝［［１５８、３３５、１４６］、～Ｘ［０］
［３３５、１４６、３７１］、～Ｘ［１］
［１４６、３７１、１０４］、～Ｘ［２］
［３７１、１０４、１７０］、～Ｘ［３］
［１０４、１７０、１０９］、～Ｘ［４］
［１７０、１０９、２９０］］～Ｘ［５］
Ｙ＝［０、０、０、０、０、０］～Ｙ［０］、Ｙ［１］、Ｙ［２］、Ｙ［３］、Ｙ［４］、Ｙ［５］
Ｓｌｉｄｉｎｇ＿Ｗｉｎｄｏｗ７＝［１０９、２９０、１２７、１５１、２３１、３７６］
Ｘ７＝［１０９、２９０、１２７］～前の過去１５分の過去３つのＣＧＭポイントに対応
Ｙ７＝０～３７６＞７０＝０、３７６ｍｇ／ｄＬ＞７０ｍｇ／ｄＬの低血糖閾値であるため、低血糖なしに対応
ここまでのＸおよびＹは以下のとおり。
Ｘ＝［［１５８、３３５、１４６］、～Ｘ［０］
［３３５、１４６、３７１］、～Ｘ［１］
［１４６、３７１、１０４］、～Ｘ［２］
［３７１、１０４、１７０］、～Ｘ［３］
［１０４、１７０、１０９］、～Ｘ［４］
［１７０、１０９、２９０］、～Ｘ［５］
［１０９、２９０、１２７］］～Ｘ［６］
Ｙ＝［０、０、０、０、０、０、０］～Ｙ［０］、Ｙ［１］、Ｙ［２］、Ｙ［３］、Ｙ［４］、Ｙ［５］、Ｙ［６］ Example of simple numerical value In the following, the above-mentioned data processing process will be described using an example of a simple numerical value. The values are randomly generated for this purpose and are not based on actual data. [KEY] Molecule: # Day: # 12 CGM values per day at mg / dL. Within the 12 CGM points of the example useful for this simplified explanation, only pHH 15 and 30 minutes ahead is possible. In the following, the calculation is mainly performed for 15 minutes pH H.
Day 0: 1: [158, 335, 146, 371, 104, 170, 109, 290, 127, 151, 231, 376]
Day 1: 2: [342, 201, 174, 100, 253, 36, 134, 270, 225, 117, 202, 356]
2: Day 3: [240, 172, 320, 174, 57, 215, 225, 163, 246, 235, 159, 36]
3: Day 4: [248, 342, 52, 388, 309, 219, 243, 275, 166, 107, 191, 288]
4: Day 5: [279, 74, 146, 276, 284, 334, 201, 185, 187, 151, 242, 114]
5: Day 6: [215, 289, 338, 282, 331, 282, 21, 152, 270, 83, 57, 114]

pHH = PH = 15 6 sliding windows.
Input: eHH of block 1:
Day 0: 1: [158, 335, 146, 371, 104, 170, 109, 290, 127, 151, 231, 376]
Sliding_Windows1 = [158, 335, 146, 371, 104, 170]
X1 = [158, 335, 146] -corresponding to the past 3 CGM points of the previous 15 minutes Y1 = 0-170> 70 = 0, 170 mg / dL> 70 mg / dL hypoglycemia threshold, so hypoglycemia Therefore, X1 will be added or added to X (or input, BG value of past CGM), and Y1 will be Y (output, binary classifier without hypoglycemia / hypoglycemia, on / off). Will be added or added to.
Sliding_Windows2 = [335, 146, 371, 104, 170, 109]
X2 = [335, 146, 371] -corresponds to the past 3 CGM points of the previous 15 minutes Y2 = 0-109> 70 = 0, 109 mg / dL> 70 mg / dL hypoglycemia threshold, so hypoglycemia Correspondence without None The X and Y up to this point are as follows.
X = [[158, 335, 146], ~ X [0]
[335, 146, 371]] to X [1]
Y = [0,0] to Ys [0], Ys [1]
Sliding_Window3 = [146, 371, 104, 170, 109, 290]
X3 = [146, 371, 104] -corresponds to the past 3 CGM points of the previous 15 minutes Y3 = 0-290> 70 = 0, 290 mg / dL> 70 mg / dL hypoglycemia threshold, so hypoglycemia Correspondence without None The X and Y up to this point are as follows.
X = [[158, 335, 146], ~ X [0]
[335, 146, 371], ~ X [1]
[146, 371, 104]] to X [2]
Y = [0, 0, 0] to Y [0], Y [1], Y [2]
Sliding_Window4 = [371, 104, 170, 109, 290, 127]
X4 = [371, 104, 170] -corresponds to the past 3 CGM points of the previous 15 minutes Y4 = 0-127> 70 = 0, 127 mg / dL> 70 mg / dL hypoglycemia threshold Correspondence without None The X and Y up to this point are as follows.
X = [[158, 335, 146], ~ X [0]
[335, 146, 371], ~ X [1]
[146, 371, 104], ~ X [2]
[371, 104, 170]] to X [3]
Y = [0,0,0,0] to Ys [0], Y [1], Y [2], Y [3]
Sliding_Window5 = [104, 170, 109, 290, 127, 151]
X5 = [104, 170, 109] -corresponds to the past 3 CGM points of the previous 15 minutes Y5 = 0 to 151> 70 = 0, 151 mg / dL> 70 mg / dL hypoglycemia threshold Corresponding to none The X and Y up to this point are as follows.
X = [[158, 335, 146], ~ X [0]
[335, 146, 371], ~ X [1]
[146, 371, 104], ~ X [2]
[371, 104, 170], ~ X [3]
[104, 170, 109]] to X [4]
Y = [0,0,0,0,0] to Y [0], Y [1], Y [2], Y [3], Y [4]
Sliding_Window6 = [170, 109, 290, 127, 151, 231]
X6 = [170, 109, 290] -corresponding to the past 3 CGM points of the previous 15 minutes Y6 = 0-231> 70 = 0, 231 mg / dL> 70 mg / dL hypoglycemia threshold Corresponding to none The X and Y up to this point are as follows.
X = [[158, 335, 146], ~ X [0]
[335, 146, 371], ~ X [1]
[146, 371, 104], ~ X [2]
[371, 104, 170], ~ X [3]
[104, 170, 109], ~ X [4]
[170, 109, 290]] to X [5]
Y = [0,0,0,0,0,0] to Y [0], Y [1], Y [2], Y [3], Y [4], Y [5]
Sliding_Window7 = [109, 290, 127, 151, 231, 376]
X7 = [109, 290, 127] -corresponding to the past 3 CGM points of the previous 15 minutes Y7 = 0-376> 70 = 0, 376 mg / dL> 70 mg / dL hypoglycemia threshold Corresponding to none The X and Y up to this point are as follows.
X = [[158, 335, 146], ~ X [0]
[335, 146, 371], ~ X [1]
[146, 371, 104], ~ X [2]
[371, 104, 170], ~ X [3]
[104, 170, 109], ~ X [4]
[170, 109, 290], ~ X [5]
[109, 290, 127]] to X [6]
Y = [0,0,0,0,0,0,0] to Y [0], Y [1], Y [2], Y [3], Y [4], Y [5], Y [ 6]

In summary, only for day 1 eHH of block 1, 7 pHH = PH = 15 Xs (inputs) with corresponding Ys (outputs) were created.
For the remaining days of eHH block 1, the values are calculated in the same way.
Day 1: 2: [342, 201, 174, 100, 253, 36, 134, 270, 225, 117, 202, 356]
2: Day 3: [240, 172, 320, 174, 57, 215, 225, 163, 246, 235, 159, 36]
The following is an example illustrating the calculations that result in the detection of hypoglycemia.
pHH = PH = 15
Day 1: 2: [342, 201, 174, 100, 253, 36, 134, 270, 225, 117, 202, 356]
Day2_Sliding_Window1 = [342, 201, 174, 100, 253, 36]
Day2_X1 = [342, 201, 174] -corresponds to the past 3 CGM points of the previous 15 minutes. Day2_Y1 = 1-36 <70 = 1, 36 mg / dL> 70 mg / dL, so hypoglycemia The X and Y up to this point are as follows.
X = [[342, 201, 174]]
Y = [1]
pHH = PH = 30
2: Day 3: [240, 172, 320, 174, 57, 215, 225, 163, 246, 235, 159, 36]
Day3_Sliding_Window1 = [240, 172, 320, 174, 57, 215, 225, 163, 246, 235, 159, 36]
Day3_X1 = [240,172,320,174,57,215] -corresponding to the past 3 CGM points of the previous 15 minutes Day3_Y1 = 1-36 <70 = 1, 36 mg / dL> 70 mg / dL hypoglycemia threshold Therefore, it corresponds to hypoglycemia. X and Y up to this point are as follows.
X = [[240, 172, 320, 174, 57, 215]]
Y = [1]

Implementation of a random forest (RF) classifier. See FIG.
The requirements for the 500 decision trees (n_estimators parameter) executed for the random forest classifier are strict. Most are performed on 100-300 decision trees. A simpler, easier-to-explain decision tree-based random forest (ANN, CNN, etc.) for the most cutting-edge, complex, but difficult-to-explain, neural networks of competitors' hypoglycemic prediction algorithms such as WellDoc, UVA, etc. It was considered reasonable to increase the number of decision trees from the more standard 100 or 300 to 500 in order to bring about the performance and competitiveness of the RF) classifier. To further fine-tune this training decision tree number parameter and other such parameters, further research and development for tolerance testing and avoiding the problem of running out of memory on local machines and local host servers, Hadoop, MapReduce, And it is necessary to move to Spark of Amazon Web Services, as well as distributed parallel computer processing using such services.

The data needs to be robust enough to adapt to such high parameters. The simply supplied raw data cannot be run on a random forest classifier with such a large number of decision trees. Therefore, innovative data preparation, transformation, adaptation, and especially optimization steps using temporal binning of rolling schemes to evaluation limit history and predictive limit history (eHH, pHH) are very important for this classification solution. In other cases, it was a solution with more guaranteed regression (although the more regressions, the more likely it is that poor data quality will occur). Classification-based solutions are much more robust and resistant to low quality data, primarily due to the data expansion and time optimization presented in the present invention.

As shown in FIG. 10, the resulting model can also be serialized in a Python object with a NumPy sequence and stored in a joblib API format that is efficient for testing different compression formats. The XZ, LZMA, and especially BZ2 formats consistently perform better (smaller size MB) compression than the Z, GZ, and especially suboptimal SAV compression formats.

To summarize the above disclosure, the use of "time binning of rolling schemes" utilizes the same amount of historical or retrospective data in a more extended, better, smarter and more fitted way. , The original raw and raw dataset can be effectively augmented and augmented.

In particular, in the evaluation limit history and predicted limit history (eHH, pHH) constructed in the process of "temporal binning of rolling schemes", random forest (RF), support vector machine (SVM), and k-nearest neighbor method (KNN). Data sets that have already been expanded are further maximized and primed to provide intervals for data that are converted and captured into ML classification methods such as.

At LTHP PH = 1 day (24 hours), RF achieved 91% accuracy, 90.9% sensitivity, and 91.9% specificity, but poor SVM and KNN performance. .. At LTHP PH = 1 day (24 hours), the SVMs performed were exacerbated with 86% accuracy, 71.4% sensitivity, and 77.4% specificity. At LTHP PH = 1 day (24 hours), KNN performed was exacerbated with 86% accuracy, 73.2% sensitivity, and 81.7% specificity. Raw CGM data was provided by Novo Nordisk clinical trial NN1218-3853.

Based on these LTHP results, only the implementation of STHP's RF was implemented in this example for the STHP ML classifier solution (named "Lombardi" in the figure). At pHH = PH = 30 minutes, the RF implementation of STHP achieved 98% accuracy, 93.59% sensitivity, and 99.75% specificity.

The STHP RF results for PH15, PH30, and PH60 are shown in FIG. FIG. 12 shows the results of the STHP RF classifier for PH15, PH30, PH45, PH60, PH75 and is compared with the results of the following published literature.

Daskalaki et al., “Real-Time Adaptive Models for the Personalized Prediction of Glycemic Profile in Type 1 Diabetes Patients.” 14 (2) 2012.
Rationale: From academic literature, the paper by Daskalaki et al. Was used as a comparison of short-term hypoglycemia predictor (STHP) classifier predictor limits (PH) at 30 and 45 minutes.

"Neural Network-Based Real-Time Prevention of Glucose in Patients with Insulin-Depend Diabetes." Diabetes Techenol, Pappada et al et al. 13 (2) 2011.
Rationale: From academic literature, the paper by Daskalaki et al. Was used as a comparison of short-term hypoglycemia predictor (STHP) classifier predictor limits (PH) in 75 minutes.

In FIGS. 13 and 14, the results of the STHP RF classifier for PH45 and PH75, respectively, are compared with the literature results. As shown, accuracy, sensitivity, and specificity were achieved at all predictive limits of 15 minutes, 30 minutes, 45 minutes, 60 minutes, and 75 minutes, which are literature comparisons from industry and academic sources. More competitive or even better than literature comparison.

Next, an example (WE) of pHH = PH = 60 minutes or STHP RF classifier 60 minutes will be described. The examples cover test code that achieves the above competitive results by loading the following five files for specific testing and validation purposes.
1. 1. STHP RF classifier model file itself: "_PH60.pkl.bz2" suffix 2. Final data of test subset of independent variable X: "_Xtest.npy" suffix 3. Final data of test subset of dependent variable y: "_ytest.npy" suffix

The following verification test metrics can be computer-processed with just the above three file inputs. Confusion matrix calculations such as raw accuracy, confusion matrix graphics themselves as well as sensitivity and specificity, as well as classification reports. See FIG.
4. Final data of all independent variables: "_X.npy" suffix 5. Final data of all dependent variables: "_y.npy" suffix

These two are only needed for cross-validated accuracy calculations. See FIG.

All of these can be combined to provide a summary report on final data entry # 1-3: WE validation test metric results: PH = 60 minutes.

See the confusion matrix table, FIG.
Calculation of confusion matrix table: TN, FN, FP, TP, see FIG.
Calculation of Confusion Matrix Table: Sensitivity, see FIG.
Calculation of Confusion Matrix Table: Specificity, see Figure 20.
Confusion matrix table calculation: Sensitivity, singularity string report output, see Figure 21.
Classification Report: Accuracy, Recall, F1 Score, and Support, see Figure 22.
For final data entry # 4 and 5: WE validation test metric results: PH = 60 minutes: Summary report: accuracy, cross-validated accuracy, sensitivity, specificity, hypoglycemic matrix (TN, FN, TP, FP) , See FIG. 23.
See the confusion matrix function, FIG. 24.
Confusion matrix function: Output (1/3), see Figure 25.
Confusion matrix function: Output (2/3): No normalization, see Figure 26.
Confusion matrix function: Output (3/3): With normalization, see Figure 27.

Cited References and Alternative Embodiments All references cited herein are individual publications, or patents, or patent applications, all incorporated by reference in their entirety for the purpose. To the extent as specifically and individually indicated, they are incorporated herein by reference in their entirety for all purposes.

All headings and subheadings are used herein for convenience only and should by no means be construed as limiting the invention.

The use of any and all examples or exemplary phrases presented herein (eg, "such as") is solely intended to make the invention more explicit and unless otherwise stated. It does not limit the scope of the invention. Nothing in the specification should be construed as indicating that any element outside the scope of the patent is essential to the practice of the invention.

The citations and incorporation of patent documents herein are for convenience only and do not reflect any aspect of the validity, patentability and / or enforceability of these patent documents.

The present invention may be implemented as a computer program product comprising a computer program mechanism embedded in a non-temporary computer readable storage medium. For example, a computer program product may include a program module shown in any combination of FIGS. 1 and 2 and / or depicted in FIG. These program modules can be stored on CD-ROMs, DVDs, magnetic disk storage products, USB keys, or any other non-temporary computer readable data or program storage products.

Many modifications and variations of the present invention can be made without departing from the spirit and scope thereof, as will be apparent to those skilled in the art. The particular embodiments described herein are provided by way of illustration only. Embodiments have been selected and described to best explain the principles of the invention and its practical applications, whereby those skilled in the art will be able to use the invention and various modifications to suit a particular application. The embodiment will be best utilized. The present invention is limited only by the terms of the appended claims and any equivalent to which such claims apply.

Claims (11)

A method for dataset optimization for improved hypoglycemia prediction based on classifier uptake, said method.
-A step of providing a raw dataset for a subject, wherein the dataset has multiple BG values acquired at a given sampling rate, and the time associated with those values over multiple days N. The process of providing, including stamps, and
-Including the step of performing data conversion by temporal binning of the rolling scheme with the evaluation block value (eHH) as the input X and creating the corresponding predicted value (pHH) as the output Y.
-X is created as a sliding window containing the BG value for a given past period T-p.
-Y is created as an indicator I indicating whether the BG value at a given future time Tf is below a given threshold indicating a hypoglycemic state. The data conversion process
-The data set optimization according to claim 1, which is performed after the step of performing data expansion by temporal binning of the daily BG value rolling scheme to the evaluation block for M days (M ≧ 2, M <N). Method for optimization. The method for dataset optimization according to claim 2, wherein the raw dataset obtained is based on an M-day insulin titration regimen. The process of providing a raw data set
-Any one of claims 1-3, performed prior to the step of performing data preparation with resampling corresponding to the nominal sampling rate and creation of an interpolated BG value to replace the missing BG value. Methods for dataset optimization as described in section. The method for data set optimization according to any one of claims 1 to 4, wherein the data conversion is performed over at least two different past time periods Tp. The method for data set optimization according to claim 5, wherein Tf corresponds to Tp. A way to train classifiers,
-A step of providing a data set optimized as defined in any one of claims 1 to 6.
-The process of importing the optimized data set into the classifier,
-A method comprising training the classifier based on the captured data set. The method for training the classifier according to claim 7, wherein the classifier is a random forest classifier. A method for predicting future BG values,
-The process of acquiring a series of evaluations of BG values from the target person,
-A step of incorporating a series of evaluations of the BG value into a classifier trained as defined in claim 7 or 8.
-A method comprising a step of providing a predicted BG value. The method for predicting future BG values according to claim 9, wherein a series of assessments of the BG values are obtained by continuous blood glucose monitoring (CGM). A computer processing system for performing temporal optimization of a data set from a subject, wherein the computer system comprises one or more processors, a memory, and the memory.
-A computer processing system comprising an instruction that, when executed by the one or more processors, implements the method defined in any one of claims 1-9.

Images